NeuralODEs for Computational Astrochemistry: The HAAS Project

In recent years, astrochemistry has seen remarkable advancements driven by cutting-edge telescopes like ALMA and JWST... However, simulating these intricate environments presents a significant challenge due to the computational demands of integrating both hydrodynamics (physics) and astrochemical models (chemistry).

The Hydrodynamical Astrochemical Autochemulators for Simulations (HAAS) project uses Neural Ordinary Differential Equations (NeuralODEs) to emulate astrochemical simulations... HAAS aims to investigate and benchmark fast and robust emulators for 3D simulations.

The project is a collaborative effort between the astrochemistry and machine learning communities... By integrating these models into hydrodynamical simulations, HAAS aspires to enable high-resolution studies that deepen our understanding of chemistry in space.

A haas is a hare in Dutch — a fitting acronym for a project that aims to accelerate astrochemistry.

Machine Learning Surrogates for Cosmological N-body Simulations

The large-scale structure of the universe is shaped by the gravitational evolution of dark matter... Cosmological N-body simulations are crucial but computationally expensive, limiting exploration of model space.

We use score-based generative machine learning models to learn the distribution of dark matter... Our approach leverages graph neural networks with message-passing layers applied to massive point cloud data.

We train the model by injecting noise and learning the score function... Sampling begins from noise and is refined iteratively using Langevin Dynamics to generate realistic structures.

Challenges include handling massive point clouds (up to 600,000 points) and periodic boundary conditions... We mitigate this using memory-efficient subgraphs and consistency-aware model design.